Sophon-Realtime-4B-Flash is a full-duplex voice dialogue model open-sourced by adventists-ai: it listens continuously while you speak, determines from your audio whether you have finished speaking, and then begins to speak while simultaneously generating responses containing sound tags such as [laughter] and [sigh], which are read aloud frame-by-frame every 80 milliseconds by a lighter-weight speech model. Officially, it can run on a single 24 GB GPU (occupying approximately 21 GB), with the first audio segment appearing in about 0.21 seconds on an RTX PRO 6000, achieving an average score of 87.7 across four metrics in Full-Duplex-Bench v1.0. It is deployed as a vLLM-Omni model, providing /v1/realtime and /v1/chat/completions interfaces; the same series also includes a more expressive Tag version and an end-to-end E2E version (both requiring 32 GB GPUs).
Why it matters The 32B version in the same series requires larger hardware; the 4B version retains features like interruptions, turn-taking, and sound tags for laughter and sighs, making it suitable for local companionship and role-playing, but its knowledge base is significantly weaker.
Hugging Face Likes 6 · Hugging Face Trending Score 5 · Hugging Face Downloads 31
License: Commercial Use Allowed · ComfyUI: Not Available · VRAM: 32 GB
Reasons
- Apache-2.0 license, downloadable from both Hugging Face and ModelScope, with an online demo space on Hugging Face
- The model card clearly details the performance gap between the 4B and 32B versions, measurement conditions, and applicable scenarios
Usage
- After git clone https://github.com/adventists-ai/sophon-realtime, enter the flash directory and run bash install.sh; then run hf download adventists-ai/Sophon-Realtime-4B-Flash --local-dir ./model, followed by GPU=0 PORT=8000 bash start.sh, and use python client_example.py --url http://127.0.0.1:8000 to converse
Risks
- Officially stated that the knowledge and reasoning capabilities of the 4B version are much weaker than those of the 32B version (VoiceBench scores 62.4 vs. 82.8); it may confidently state incorrect facts, so use the 32B version when answer accuracy is required
- It processes only voice and has no visual capabilities; it supports zero-shot voice cloning but can only clone your own voice or voices for which authorization has been obtained; the hosted API recommended at the beginning of the model card is an official paid service